Reddit thread: agents can manufacture genuine evidence by acting on unverified claims

usefulwrong · reddit · 2026-11-17

Using ProPublica's investigation of Brownsville Energy Storage Terminals as a starting point, the author shows how Google AI Overview already repeats unverifiable claims from a shell company's website.

He then describes a subtler failure mode in multi-agent systems: one agent acts on unverified info (creates a vendor record, opens an inquiry); months later another agent treats those genuinely existing records as confirmation of the original claim. Since the new records have their own origins, source-tracing won't necessarily catch this. He asks how builders track not just where information came from, but which decisions exist because an earlier system believed it.

Original post →

More from AGI Musings

AGI Musings channel →